Computation and Language
AutoHarness: improving LLM agents by automatically synthesizing a code harness
Xinghua Lou, Miguel Lázaro-Gredilla, Antoine Dedieu et al. (+3 more)
Despite significant strides in
language models in the last few years, when used as
agents, such models often try to perform actions that are not just suboptimal for a given state, but are strictly prohibited by the external environment. For example, in the recent Kaggle GameArena chess competition, 78% of Gemini-2.5-Flash losses were attributed to illegal moves. Often people manually write "harnesses" around LLMs to prevent such failures. In this paper, we demonstrate that Gemini-2.5-Flash can automatically synthesize such a code harness, using a small number of rounds of iterative code refinement given feedback from the (game) environment. The resulting harness prevents all illegal moves in 145 different TextArena games (both 1-player and 2-player), enabling the smaller Gemini-2.5-Flash model to outperform larger models, such as Gemini-2.5-Pro. Pushing our technique to the limit, we can get Gemini-2.5-Flash to generate the entire policy in code, thus eliminating the need to use the LLM at decision making time. The resulting code-policy receives a higher average reward than Gemini-2.5-Pro and GPT-5.2-High on 16 TextArena 1-player games. Our results show that using a smaller model to synthesize a custom code harness (or entire policy) can outperform a much larger model, while also being more cost effective.
Read the full abstract ▾
All authors Xinghua Lou, Miguel Lázaro-Gredilla, Antoine Dedieu, Carter Wendelken, Wolfgang Lehrach, Kevin P. Murphy
Computation and Language
StructLens: A Structural Lens for Language Models via Maximum Spanning Trees
Haruki Sakajo, Frederikus Hudi, Yusuke Sakai et al. (+2 more)
Language exhibits inherent structures, a property that explains both
language acquisition and
language change. Given this characteristic, we expect
language models to manifest their own internal structures as well. While interpretability research has investigated how models compute representations mechanistically through
attention patterns and Sparse AutoEncoders, the organization of the resulting representations is overlooked. To address this gap, we introduce StructLens, a framework to analyze representations through a holistic structural view. StructLens constructs maximum spanning trees based on the semantic representations in residual streams, inspired by tree representation in dependency parsing, and provides summaries of
token relationships in representation space. We analyze how contiguous
tokens are also nearby in representation space and find that middle layers show the strongest local-span organization. Moreover, analysis of pre-training checkpoints reveals that smaller local units become detectable earlier in pre-training, and larger units later. Our findings demonstrate that StructLens provides insights into how models organize
token representations across layers and training. Our code is available at https://github.com/naist-nlp/structlens.
Read the full abstract ▾
All authors Haruki Sakajo, Frederikus Hudi, Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe
Computation and Language
A benchmark for joint dialogue satisfaction, emotion recognition, and emotion state transition prediction
Jing Bian, Haoxiang Su, Liting Jiang et al. (+6 more)
User satisfaction is closely related to enterprises, as it not only directly reflects users' subjective evaluation of service quality or products, but also affects customer loyalty and long-term business revenue. Monitoring and understanding user
emotions during interactions helps predict and improve satisfaction. However, relevant Chinese
datasets are limited, and user
emotions are dynamic; relying on single-turn dialogue cannot fully track emotional changes across multiple turns, which may affect satisfaction prediction. To address this, we constructed a multi-task, multi-label Chinese dialogue
dataset that supports satisfaction recognition, as well as
emotion recognition and emotional state transition prediction, providing new resources for studying
emotion and satisfaction in dialogue systems.
Read the full abstract ▾
All authors Jing Bian, Haoxiang Su, Liting Jiang, Di Wu, Ruiyu Fang, Xiaomeng Huang, Yanbing Li, Shuangyong Song, Hao Huang
Computation and Language
Controllable and explainable personality sliders for LLMs at inference time
Florian Hoppe, David Khachaturov, Robert Mullins et al. (+1 more)
Aligning Large
Language Models (LLMs) with specific personas typically relies on expensive and monolithic Supervised Fine-Tuning (SFT) or RLHF. While effective, these methods require training distinct models for every target personality profile. Inference-time activation steering offers a parameter-efficient alternative, yet naive approaches fail to control multiple traits simultaneously due to destructive vector interference. In this work, we propose a modular framework for continuous, multi-dimensional personality control. Our key innovation is Sequential Adaptive Steering (SAS): a method that orthogonalizes steering vectors by training subsequent probes on the residual stream shifted by prior interventions. This approach transforms steering vectors into reusable primitives, allowing users to instantly synthesize complex, high-fidelity personality profiles by simply adjusting coefficients alpha. We validate our framework on the Big Five personality traits, demonstrating that it outperforms naive baselines in both goal adherence and coherence, enabling precise, holistic personality modulation without updating model parameters.
Read the full abstract ▾
All authors Florian Hoppe, David Khachaturov, Robert Mullins, Mark Huasong Meng
Computation and Language
IntPro: A Proxy Agent for Context-Aware Intent Understanding via Retrieval-conditioned Inference
Guanming Liu, Meng Wu, Peng Zhang et al. (+8 more)
Large
language models (LLMs) have become integral to modern Human-AI collaboration workflows, where accurately understanding user intent serves as a crucial step for generating satisfactory responses. Context-aware intent understanding, which involves inferring user intentions from situational environments, is inherently challenging because it requires
reasoning over both the immediate context and the user's underlying motivations that drive their behavior. Moreover, existing approaches often treat intent understanding as a static recognition task, overlooking users' accumulated intent patterns that could provide valuable references for more accurate and generalizable understanding. To address this gap, we propose IntPro, a proxy
agent that learns to adapt to individual users via
retrieval-conditioned intent inference. We design intent explanations that abstract how contextual signals connect to expressed intents, and store them in an individual intent history library for
retrieval. We train IntPro through supervised fine-tuning on
retrieval-conditioned trajectories and multi-turn Group Relative Policy Optimization (GRPO) with tool-aware reward functions, enabling the
agent to learn when to leverage historical intent patterns and when to infer directly. Experiments across three diverse scenarios (Highlight-Intent, MIntRec2.0, and Weibo Post-Sync) demonstrate that IntPro achieves strong intent understanding performance with effective context-aware
reasoning capabilities across different scenarios and model types.
Read the full abstract ▾
All authors Guanming Liu, Meng Wu, Peng Zhang, Yu Zhang, Yubo Shu, Xianliang Huang, Kainan Tu, Ning Gu, Liuxin Zhang, Qianying Wang, Tun Lu
Computation and Language
Controlling Chat Style in Language Models via Single-Direction Editing
Zhenyu Xu and Victor S. Sheng
Controlling stylistic attributes in large
language models (LLMs) remains challenging, with existing approaches relying on either
prompt engineering or post-training
alignment. This paper investigates this challenge through the lens of representation engineering, testing the hypothesis that distinct stylistic attributes - from emotional tone to linguistic structure - are encoded as linear directions in the model's activation space. We provide strong empirical evidence for this hypothesis across a wide range of styles and, based on this finding, present a lightweight, training-free method for precise style control. Our approach supports linear style composition, enhances safety by ablating undesirable behaviors, and, as confirmed by experiments on over a dozen models, achieves high style adherence while preserving core capabilities at minimal computational cost.
Read the full abstract ▾
Computation and Language
Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinement
Yuxiao Lu, Lin Xu, Yang Sun et al. (+2 more)
Large
language models (LLMs) aligned for safety often suffer from over-refusal, the tendency to reject seemingly toxic or benign
prompts by misclassifying them as toxic. This behavior undermines models' helpfulness and restricts usability in sensitive or nuanced contexts. While prior work has proposed mitigation strategies such as data augmentation and activation steering, these approaches often face a trade-off: reducing over-refusal typically degrades the model's ability to reject genuinely harmful content. We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic
prompts on the model's learning dynamics. To address it, we introduce a preceding
alignment stage, DCR: Discernment via Contrastive Refinement. Both theoretically and empirically, we demonstrate that contrastive refinement improves an LLM's capacity to distinguish truly toxic
prompts from superficially toxic ones. Evaluation across diverse
benchmarks shows that our method effectively reduces over-refusal while preserving the safety benefits of
alignment. Importantly, it achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety
alignment.
Read the full abstract ▾
All authors Yuxiao Lu, Lin Xu, Yang Sun, Wenjun Li, Jie Shi
Computation and Language
Can Large Language Models Derive New Knowledge? A Dynamic Benchmark for Biological Knowledge Discovery
Chaoqun Yang, Xinyu Lin, Shulin Li et al. (+4 more)
Recent advancements in Large
Language Model (LLM)
agents have demonstrated remarkable potential in automatic knowledge discovery. However, rigorously evaluating an AI's capacity for knowledge discovery remains a critical challenge. Existing
benchmarks predominantly rely on static
datasets, leading to inevitable data contamination where models have likely seen the evaluation knowledge during training. Furthermore, the rapid release cycles of modern LLMs render static
benchmarks quickly outdated, failing to assess the ability to discover truly new knowledge. To address these limitations, we propose DBench-Bio, a dynamic and fully automated
benchmark designed to evaluate AI's biological knowledge discovery ability. DBench-Bio employs a three-stage pipeline: (1) data acquisition of rigorous, authoritative paper abstracts; (2) QA extraction utilizing LLMs to synthesize scientific hypothesis questions and corresponding discovery answers; and (3) QA filter to ensure quality based on relevance, clarity, and centrality. We instantiate this pipeline to construct a monthly-updated
benchmark covering 12 biomedical sub-domains. Extensive evaluations of SOTA models reveal current limitations in discovering new knowledge. Our work provides the first dynamic, automatic framework for assessing the new knowledge discovery capabilities of AI systems, establishing a living, evolving resource for AI research community to catalyze the development of knowledge discovery.
Read the full abstract ▾
All authors Chaoqun Yang, Xinyu Lin, Shulin Li, Wenjie Wang, Ruihan Guo, Fuli Feng, Tat-Seng Chua
Computation and Language
DIALEVAL: Automated Type-Theoretic Evaluation of LLM Instruction Following
Nardine Basta and Dali Kaafar
Evaluating instruction following in Large
Language Models requires decomposing instructions into verifiable requirements and assessing satisfaction–tasks currently dependent on manual annotation and uniform criteria that do not align with human judgment patterns. We present DIALEVAL, a type-theoretic framework using dual LLM
agents to automate instruction decomposition into typed predicates and implement type-specific satisfaction semantics. The framework enforces formal atomicity and independence constraints during automated extraction, then applies differentiated evaluation criteria–semantic equivalence for content predicates, exact precision for numerical predicates–mirroring empirically observed human assessment patterns. Extended to multi-turn dialogues through history-aware satisfaction functions, DIALEVAL enables evaluation in conversational contexts where single-turn methods fail. Validation demonstrates 90.38% accuracy (26.45% error reduction over baselines) and substantially stronger correlation with human judgment for complex instructions.
Read the full abstract ▾
Computation and Language
From We to Me: Theory Informed Narrative Shift with Abductive Reasoning
Jaikrishna Manojkumar Patil, Divyagna Bavikadi, Kaustuv Mukherji et al. (+5 more)
Effective communication often relies on aligning a message with an audience's narrative and worldview. Narrative shift involves transforming text to reflect a different narrative framework while preserving its original core message–a task we demonstrate is significantly challenging for current Large
Language Models (LLMs). To address this, we propose a neurosymbolic approach grounded in social science theory and abductive
reasoning. Our method automatically extracts rules to abduce the specific story elements needed to guide an LLM through a consistent and targeted narrative transformation. Across multiple LLMs, abduction-guided transformed stories shifted the narrative while maintaining the fidelity with the original story. For example, with GPT-4o we outperform the zero-shot LLM baseline by 55.88% for collectivistic to individualistic narrative shift while maintaining superior semantic similarity with the original stories (40.4% improvement in KL divergence). For individualistic to collectivistic transformation, we achieve comparable improvements. We show similar performance across both directions for Llama-4, and Grok-4 and competitive performance for Deepseek-R1.
Read the full abstract ▾
All authors Jaikrishna Manojkumar Patil, Divyagna Bavikadi, Kaustuv Mukherji, Ashby Steward-Nolan, Peggy-Jean Allin, Tumininu Awonuga, Joshua Garland, Paulo Shakarian
Computation and Language · Cross-listed from cs.LG
Self-Play Only Evolves When Self-Synthetic Pipeline Ensures Learnable Information Gain
Wei Liu, Siya Qi, Yali Du et al. (+1 more)
Large
language models (LLMs) make it plausible to build systems that improve through self-evolving loops, but many existing proposals are better understood as self-play and often plateau quickly. A central failure mode is that the loop synthesises more data without increasing learnable information for the next iteration. Through experiments on a self-play coding task, we reveal that sustainable self-
evolution requires a self-synthesised data pipeline with learnable information that increases across iterations. We identify triadic roles that self-evolving LLMs play: the Proposer, which generates tasks; the Solver, which attempts solutions; and the Verifier, which provides training signals, and we identify three system designs that jointly target learnable information gain from this triadic roles perspective. Asymmetric co-
evolution closes a weak-to-strong-to-weak loop across roles. Capacity growth expands parameter and inference-time budgets to match rising learnable information. Proactive information seeking introduces external context and new task sources that prevent saturation. Together, these modules provide a measurable, system-level path from brittle self-play dynamics to sustained self-
evolution.
Read the full abstract ▾
All authors Wei Liu, Siya Qi, Yali Du, Yulan He
Computation and Language · Cross-listed from cs.CY
How effective are VLMs in assisting humans in inferring the quality of mental models from Multimodal short answers?
Pritam Sil, Durgaprasad Karnam, Vinay Reddy Venumuddala et al. (+1 more)
STEM Mental models can play a critical role in assessing students' conceptual understanding of a topic. They not only offer insights into what students know but also into how effectively they can apply, relate to, and integrate concepts across various contexts. Thus, students' responses are critical markers of the quality of their understanding and not entities that should be merely graded. However, inferring these mental models from student answers is challenging as it requires deep
reasoning skills. We propose MMGrader, an approach that infers the quality of students' mental models from their
multimodal responses using concept graphs as an analytical framework. In our evaluation with 9 openly available models, we found that the best-performing models fall short of human-level performance. This is because they only achieved an accuracy of approximately 40%, a prediction error of 1.1 units, and a scoring distribution fairly aligned with human scoring patterns. With improved accuracy, these can be highly effective assistants to teachers in inferring the mental models of their entire classrooms, enabling them to do so efficiently and help improve their pedagogies more effectively by designing targeted help sessions and lectures that strengthen areas where students collectively demonstrate lower proficiency.
Read the full abstract ▾
All authors Pritam Sil, Durgaprasad Karnam, Vinay Reddy Venumuddala, Pushpak Bhattacharyya
Computation and Language · Cross-listed from cs.CY
Red Teaming LLMs as Socio-Technical Practice: From Exploration and Data Creation to Evaluation
Adriana Alvarado Garcia, Ruyuan Wan, Ozioma C. Oguine et al. (+1 more)
Recently, red teaming, with roots in security, has become a key evaluative approach to ensure the safety and reliability of Generative Artificial Intelligence. However, most existing work emphasizes technical
benchmarks and attack success rates, leaving the socio-technical practices of how red teaming
datasets are defined, created, and evaluated under-examined. Drawing on 22 interviews with practitioners who design and evaluate red teaming
datasets, we examine the data practices and standards that underpin this work. Because
adversarial datasets determine the scope and accuracy of model evaluations, they are critical artifacts for assessing potential harms from large
language models. Our contributions are first, empirical evidence of practitioners conceptualizing red teaming and developing and evaluating red teaming
datasets. Second, we reflect on how practitioners' conceptualization of risk leads to overlooking the context, interaction type, and user specificity. We conclude with three opportunities for HCI researchers to expand the conceptualization and data practices for red-teaming.
Read the full abstract ▾
All authors Adriana Alvarado Garcia, Ruyuan Wan, Ozioma C. Oguine, Karla Badillo-Urquiola
Computation and Language · Cross-listed from cs.LG
Hardware Co-Design Scaling Laws via Roofline Modelling for On-Device LLMs
Luoyang Sun, Jiwen Jiang, Yifeng Ding et al. (+9 more)
Vision-
Language-Action Models (VLAs) have emerged as a key paradigm of Physical AI and are increasingly deployed in autonomous vehicles,
robots, and smart spaces. In these resource-constrained on-device settings, selecting an appropriate large
language model (LLM) backbone is a critical challenge: models must balance accuracy with strict inference latency and hardware efficiency constraints. This makes hardware-software co-design a game-changing requirement for on-device LLM deployment, where each hardware platform demands a tailored architectural solution. We propose a hardware co-design law that jointly captures model accuracy and inference performance. Specifically, we model training loss as an explicit function of architectural hyperparameters and characterise inference latency via roofline modelling. We empirically evaluate 1,942 candidate architectures on NVIDIA Jetson Orin, training 170 selected models for 10B
tokens each to fit a scaling law relating architecture to training loss. By coupling this scaling law with latency modelling, we establish a direct accuracy-latency correspondence and identify the Pareto frontier for hardware co-designed LLMs. We further formulate architecture search as a joint optimisation over precision and performance, deriving feasible design regions under industrial hardware and application budgets. Our approach reduces architecture selection from months to days. At the same latency as Qwen2.5-0.5B on the target hardware, our co-designed architecture achieves 19.42% lower perplexity on WikiText-2. To our knowledge, this is the first principled and operational framework for hardware co-design scaling laws in on-device LLM deployment. We will make the code and related checkpoints publicly available.
Read the full abstract ▾
All authors Luoyang Sun, Jiwen Jiang, Yifeng Ding, Fengfa Li, Yan Song, Haifeng Zhang, Jian Ying, Lei Ren, Kun Zhan, Wei Chen, Yan Xie, Cheng Deng
Computation and Language
Autonomous Continual Learning for Environment Adaptation of Computer-Use Agents
Tianci Xue, Zeyi Liao, Tianneng Shi et al. (+5 more)
Real-world digital environments are highly diverse and dynamic. These characteristics cause
agents to frequently encounter unseen environments and distribution shifts, making continual learning in such environments essential for computer-use
agents (CUAs). However, a key challenge lies in obtaining high-quality and environment-grounded training data without relying on costly human annotation. In this work, we introduce ACuRL, an Autonomous Curriculum Reinforcement Learning framework that continually adapts
agents to specific environments with zero human data. The
agent first explores an environment to acquire initial experiences. During subsequent iterative training, a curriculum task generator leverages these experiences together with feedback from the previous iteration to synthesize new tasks tailored for the
agent's current capabilities. To provide reliable reward signals, we introduce CUAJudge, a robust automatic evaluator for CUAs that achieves 93% agreement with human judgments. Empirically, our method effectively enables both intra-environment and cross-environment continual learning, yielding 3-29% absolute performance gains on the target environments without catastrophic forgetting on others. We also show that it can mitigate performance degradation under environment changes (e.g., version updates, platform migration, and resolution shifts). Further analyses show highly sparse updates (e.g., only 20% parameters), which helps explain the effective and robust adaptation.
Read the full abstract ▾
All authors Tianci Xue, Zeyi Liao, Tianneng Shi, Zilu Wang, Kai Zhang, Dawn Song, Yu Su, Huan Sun
Computation and Language
Physically Interpretable AlphaEarth Foundation Model Embeddings Enable LLM-Based Land Surface Intelligence
Mashrekur Rahman
Satellite foundation models produce dense
embeddings whose physical interpretability remains poorly understood, limiting their integration into environmental decision systems. Using 12.1 million samples across the Continental United States (2017–2023), we first present a comprehensive interpretability analysis of Google AlphaEarth's 64-dimensional
embeddings against 26 environmental variables spanning
climate, vegetation, hydrology, temperature, and terrain. Combining linear, nonlinear, and
attention-based methods, we show that individual
embedding dimensions map onto specific land surface properties, while the full
embedding space reconstructs most environmental variables with high fidelity (12 of 26 variables exceed $R^2 > 0.90$; temperature and elevation approach $R^2 = 0.97$). The strongest dimension-variable relationships converge across all three analytical methods and remain robust under spatial block cross-validation (mean $\Delta R^2 = 0.017$) and temporally stable across all seven study years (mean inter-year correlation $r = 0.963$). Building on these validated interpretations, we then developed a Land Surface Intelligence system that implements
retrieval-augmented generation over a FAISS-indexed
embedding database of 12.1 million vectors, translating natural
language environmental queries into satellite-grounded assessments. An LLM-as-Judge evaluation across 360 query–response cycles, using four LLMs in rotating generator, system, and judge roles, achieved weighted scores of $\mu = 3.74 \pm 0.77$ (scale 1–5), with grounding ($\mu = 3.93$) and coherence ($\mu = 4.25$) as the strongest criteria. Our results demonstrate that satellite foundation model
embeddings are physically structured representations that can be operationalized for environmental and geospatial intelligence.
Read the full abstract ▾
Computation and Language
Learning Self-Interpretation from Interpretability Artifacts: Training Lightweight Adapters on Vector-Label Pairs
Keenan Pepper, Alex McKenzie, Florin Pop et al. (+6 more)
Self-interpretation methods
prompt language models to describe their own internal states, but remain unreliable due to hyperparameter sensitivity. We show that training lightweight adapters on interpretability artifacts, while keeping the LM entirely frozen, yields reliable self-interpretation across tasks and model families. A scalar affine adapter with just $d_\text{model}+1$ parameters suffices: trained adapters generate sparse autoencoder feature labels that outperform the training labels themselves (70% vs 50% generation scoring at 70B scale), identify topics with 94% recall@1 versus 1% for untrained baselines, and decode bridge entities in multi-hop
reasoning that appear in neither
prompt nor response, surfacing implicit
reasoning without chain-of-thought. The learned bias vector alone accounts for 85% of improvement, and simpler adapters generalize better than more expressive alternatives. Controlling for model knowledge via prompted descriptions, we find self-interpretation gains outpace capability gains from 7B to 72B parameters. Our results demonstrate that self-interpretation improves with scale, without modifying the model being interpreted.
Read the full abstract ▾
All authors Keenan Pepper, Alex McKenzie, Florin Pop, Stijn Servaes, Martin Leitgab, Mike Vaiana, Judd Rosenblatt, Michael S. A. Graziano, Diogo de Lucena
Computation and Language
When Less Is More? Diagnosing ASR Predictions in Sardinian via Layer-Wise Decoding
Domenico De Cristofaro, Alessandro Vietti, Marianne Pouplier et al. (+1 more)
Recent studies have shown that intermediate layers in
multilingual speech models often encode more phonetically accurate representations than the final output layer. In this work, we apply a layer-wise decoding strategy to a pretrained Wav2Vec2 model to investigate how phoneme-level predictions evolve across encoder layers, focusing on Campidanese Sardinian, a low-resource
language. We show that truncating upper
transformer layers leads to improved Phoneme Error Rates (PER), with the best performance achieved not at the final layer, but two layers earlier. Through fine-grained
alignment analysis, we find that intermediate predictions better preserve segmental identity, avoid overgeneration, and reduce certain classes of phonological errors. We also introduce the notion of regressive errors, cases where correct predictions at intermediate layers are overwritten by errors at the final layer. These regressions highlight the limitations of surface-level error metrics and reveal how deeper layers may generalize or abstract away from acoustic detail. Our findings support the use of early-layer probing as a diagnostic tool for ASR models, particularly in low-resource settings where standard evaluation metrics may fail to capture linguistically meaningful behavior.
Read the full abstract ▾
All authors Domenico De Cristofaro, Alessandro Vietti, Marianne Pouplier, Aleese Block
Computation and Language
Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models
Arash Gholami Davoodi, Navid Rezazadeh, Seyed Pouyan Mousavi Davoudi et al. (+1 more)
Large
language models (LLMs) must balance diversity and creativity against logical coherence in open-ended generation. Existing truncation-based samplers are effective but largely heuristic, relying mainly on probability mass and
entropy while ignoring semantic geometry of the
token space. We present Top-W, a geometry-aware truncation rule that uses Wasserstein distance-defined over
token-
embedding geometry-to keep the cropped distribution close to the original, while explicitly balancing retained probability mass against the
entropy of the kept set. Our theory yields a simple closed-form structure for the fixed-potential subset update: depending on the mass-
entropy trade-off, the optimal crop either collapses to a single
token or takes the form of a one-dimensional prefix that can be found efficiently with a linear scan. We implement Top-W using efficient geometry-based potentials (nearest-set or k-NN) and pair it with an alternating decoding routine that keeps the standard truncation-and-sampling interface unchanged. Extensive experiments on four
benchmarks (GSM8K, GPQA, AlpacaEval, and MT-Bench) across three instruction-tuned models show that Top-W consistently outperforms prior state-of-the-art decoding approaches achieving up to 33.7% improvement. Moreover, we find that Top-W not only improves accuracy-focused performance, but also boosts creativity under judge-based open-ended evaluation.
Read the full abstract ▾
All authors Arash Gholami Davoodi, Navid Rezazadeh, Seyed Pouyan Mousavi Davoudi, Pouya Pezeshkpour
Computation and Language
The Subjectivity of Respect in Police Traffic Stops: Modeling Community Perspectives in Body-Worn Camera Footage
Preni Golazizian, Elnaz Rahmati, Jackson Trager et al. (+17 more)
Traffic stops are among the most frequent police-civilian interactions, and body-worn cameras (BWCs) provide a unique record of how these encounters unfold. Respect is a central dimension of these interactions, shaping public trust and perceived legitimacy, yet its interpretation is inherently subjective and shaped by lived experience, rendering community-specific perspectives a critical consideration. Leveraging unprecedented access to Los Angeles Police Department BWC footage, we introduce the first large-scale traffic-stop
dataset annotated with respect ratings and free-text rationales from multiple perspectives. By sampling annotators from police-affiliated, justice-system-impacted, and non-affiliated Los Angeles residents, we enable the systematic study of perceptual differences across diverse communities. To this end, we (i) develop a domain-specific evaluation rubric grounded in procedural justice theory, LAPD training materials, and extensive fieldwork; (ii) introduce a rubric-driven preference data construction framework for perspective-consistent
alignment; and (iii) propose a perspective-aware modeling framework that predicts personalized respect ratings and generates annotator-specific rationales for both officers and civilian drivers from traffic-stop transcripts. Across all three annotator groups, our approach improves both rating prediction performance and rationale
alignment. Our perspective-aware framework enables law enforcement to better understand diverse community expectations, providing a vital tool for building public trust and procedural legitimacy.
Read the full abstract ▾
All authors Preni Golazizian, Elnaz Rahmati, Jackson Trager, Zhivar Sourati, Nona Ghazizadeh, Georgios Chochlakis, Jose Alcocer, Kerby Bennett, Aarya Vijay Devnani, Parsa Hejabi, Harry G. Muttram, Akshay Kiran Padte, Mehrshad Saadatinia, Chenhao Wu, Alireza S. Ziabari, Michael Sierra-Arévalo, Nick Weller, Shrikanth Narayanan, Benjamin A. T. Graham, Morteza Dehghani
Computation and Language
Are More Tokens Rational? Inference-Time Scaling in Language Models as Adaptive Resource Rationality
Zhimin Hu, Riya Roshan and Sashank Varma
Human
reasoning is shaped by resource rationality – optimizing performance under constraints. Recently, inference-time scaling has emerged as a powerful paradigm to improve the
reasoning performance of Large
Language Models by expanding test-time computation. Specifically, instruction-tuned (IT) models explicitly generate long
reasoning steps during inference, whereas Large
Reasoning Models (LRMs) are trained by reinforcement learning to discover
reasoning paths that maximize accuracy. However, it remains unclear whether resource-rationality can emerge from such scaling without explicit reward related to computational costs. We introduce a Variable Attribution Task in which models infer which variables determine outcomes given candidate variables, input-output trials, and predefined logical functions. By varying the number of candidate variables and trials, we systematically manipulate task complexity. Both models exhibit a transition from brute-force to analytic strategies as complexity increases. IT models degrade on XOR and XNOR functions, whereas LRMs remain robust. These findings suggest that models can adjust their
reasoning behavior in response to task complexity, even without explicit cost-based reward. It provides compelling evidence that resource rationality is an emergent property of inference-time scaling itself.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Discovering Differences in Strategic Behavior Between Humans and LLMs
Caroline Wang, Daniel Kasenberg, Kim Stachenfeld et al. (+1 more)
As Large
Language Models (LLMs) are increasingly deployed in social and strategic scenarios, it becomes critical to understand where and why their behavior diverges from that of humans. While behavioral game theory (BGT) provides a framework for analyzing behavior, existing models do not fully capture the idiosyncratic behavior of humans or black-box, non-human
agents like LLMs. We employ AlphaEvolve, a cutting-edge program discovery tool, to directly discover interpretable models of human and LLM behavior from data, thereby enabling open-ended discovery of structural factors driving human and LLM behavior. Our analysis on iterated rock-paper-scissors reveals that frontier LLMs can be capable of deeper strategic behavior than humans. These results provide a foundation for understanding structural differences driving differences in human and LLM behavior in strategic interactions.
Read the full abstract ▾
All authors Caroline Wang, Daniel Kasenberg, Kim Stachenfeld, Pablo Samuel Castro
Computation and Language
On Emergent Social World Models – Evidence for Functional Integration of Theory of Mind and Pragmatic Reasoning in Language Models
Polina Tsvilodub, Jan-Felix Klumpp, Amir Mohammadpour et al. (+2 more)
This paper investigates whether LMs recruit shared computational mechanisms for general Theory of Mind (ToM) and
language-specific pragmatic
reasoning in order to contribute to the general question of whether LMs may be said to have emergent "social world models", i.e., representations of mental states that are repurposed across tasks (the functional integration hypothesis). Using behavioral evaluations and causal-mechanistic experiments via functional localization methods inspired by cognitive neuroscience, we analyze LMs' performance across seven subcategories of ToM abilities (Beaudoin et al., 2020) on a substantially larger localizer
dataset than used in prior like-minded work. Results from stringent hypothesis-driven statistical testing offer suggestive evidence for the functional integration hypothesis, indicating that LMs may develop interconnected "social world models" rather than isolated competencies. This work contributes novel ToM localizer data, methodological refinements to functional localization techniques, and empirical insights into the emergence of social cognition in artificial systems.
Read the full abstract ▾
All authors Polina Tsvilodub, Jan-Felix Klumpp, Amir Mohammadpour, Jennifer Hu, Michael Franke
Computation and Language · Cross-listed from cs.IR
MLDocRAG: Multimodal Long-Context Document Retrieval Augmented Generation
Yongyue Zhang and Yaxiong Wu
Understanding
multimodal long-context documents that comprise
multimodal chunks such as paragraphs, figures, and tables is challenging due to (1) cross-modal heterogeneity to localize relevant information across modalities, (2) cross-page
reasoning to aggregate dispersed evidence across pages. To address these challenges, we are motivated to adopt a query-centric formulation that projects cross-modal and cross-page information into a unified query representation space, with queries acting as abstract semantic surrogates for heterogeneous
multimodal content. In this paper, we propose a
Multimodal Long-Context Document
Retrieval Augmented Generation (MLDocRAG) framework that leverages a
Multimodal Chunk-Query Graph (MCQG) to organize
multimodal document content around semantically rich, answerable queries. MCQG is constructed via a
multimodal document expansion process that generates fine-grained queries from heterogeneous document chunks and links them to their corresponding content across modalities and pages. This graph-based structure enables selective, query-centric
retrieval and structured evidence aggregation, thereby enhancing grounding and coherence in
multimodal long-context question answering. Experiments on
datasets MMLongBench-Doc and LongDocURL demonstrate that MLDocRAG consistently improves
retrieval quality and answer accuracy, demonstrating its effectiveness for
multimodal long-context understanding.
Read the full abstract ▾
Computation and Language
Learning to Evict from Key-Value Cache
Luca Moschella, Laura Manduchi and Ozan Sener
The growing size of Large
Language Models (LLMs) makes efficient inference challenging, primarily due to the
memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past
attention scores, which serve only as indirect proxies for a
token's future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank
tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of lightweight per-head RL
agents trained on pre-computed generation traces using only key and value vectors. Each
agent learns a specialized eviction policy guided by a holistic reward, derived from future utility, that evaluates the quality of the ranking across all cache budgets, requiring no modifications to the underlying LLM or additional inference. Evaluated across two model families on the long-context
benchmark RULER (up to 128K
tokens) and the multi-turn dialogue
benchmark OASST2-4k, KVP significantly outperforms strong baselines. Zero-shot tests on standard downstream tasks (BoolQ, LongBench passage
retrieval, GovReport) further show that KVP generalizes beyond its training distribution and to considerably longer sequence lengths. These results demonstrate that learning to predict future
token utility is a powerful and scalable paradigm for adaptive KV cache management.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Blockwise Advantage Estimation for Multi-Objective RL with Verifiable Rewards
Kirill Pavlenko, Alexander Golubev, Simon Karasik et al. (+1 more)
Group Relative Policy Optimization (GRPO) assigns a single scalar advantage to all
tokens in a completion. For structured generations with explicit segments and objectives, this couples unrelated reward signals across segments, leading to objective interference and misattributed credit. We propose Blockwise Advantage Estimation, a family of GRPO-compatible methods that assigns each objective its own advantage and applies it only to the
tokens in the corresponding text block, reducing reliance on hand-designed scalar rewards and scaling naturally to additional objectives. A key challenge is estimating advantages for later blocks whose rewards are conditioned on sampled prefixes; standard unbiased approaches require expensive nested rollouts from intermediate states. Concretely, we introduce an Outcome-Conditioned Baseline that approximates intermediate state values using only within-group statistics by stratifying samples according to a prefix-derived intermediate outcome. On math tasks with uncertainty estimation, our method mitigates reward interference, is competitive with a state-of-the-art reward-designed approach, and preserves test-time gains from confidence-weighted ensembling. More broadly, it provides a modular recipe for optimizing sequential objectives in structured generations without additional rollouts.
Read the full abstract ▾
All authors Kirill Pavlenko, Alexander Golubev, Simon Karasik, Boris Yangel
Computation and Language
Latent Thoughts Tuning: Bridging Context and Reasoning with Fused Information in Latent Tokens
Weihao Liu, Dehai Min and Lu Cheng
While explicit Chain-of-Thought (CoT) equips Large
Language Models (LLMs) with strong
reasoning capabilities, it constrains the model's thoughts to a discrete vocabulary space. Recently,
reasoning in continuous latent space has emerged as a promising alternative, but current paradigms suffer from feature collapse and instability due to distribution mismatch when recurrently reusing hidden states, or
alignment issues when relying on assistant models. To address this, we propose Latent Thoughts Tuning (LT-Tuning), a post-training framework that redefines how latent thoughts are constructed and deployed. Instead of relying solely on raw hidden states, our method introduces a Context-Prediction-Fusion mechanism that jointly leverages contextual hidden states and predictive semantic guidance from the vocabulary
embedding space. Combined with a progressive three-stage curriculum learning pipeline, LT-Tuning also enables dynamic switching between latent and explicit thinking modes. Experiments demonstrate that our method outperforms existing latent
reasoning baselines, effectively mitigating feature collapse and achieving robust
reasoning accuracy.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Towards Autonomous Mathematics Research
Tony Feng, Trieu H. Trinh, Garrett Bingham et al. (+25 more)
Recent advances in foundational models have yielded
reasoning systems capable of achieving a gold-medal standard at the International Mathematical Olympiad. The transition from competition-level problem-solving to professional research, however, requires navigating vast literature and constructing long-horizon proofs. In this work, we introduce Aletheia, a math research
agent that iteratively generates, verifies, and revises solutions end-to-end in natural
language. Specifically, Aletheia is powered by an advanced version of Gemini Deep Think for challenging
reasoning problems, a novel inference-time scaling law that extends beyond Olympiad-level problems, and intensive tool use to navigate the complexities of mathematical research. We demonstrate the capability of Aletheia from Olympiad problems to PhD-level exercises and most notably, through several distinct milestones in AI-assisted mathematics research: (a) a research paper (Feng26) generated by AI without any human intervention in calculating certain structure constants in arithmetic geometry called eigenweights; (b) a research paper (LeeSeo26) demonstrating human-AI collaboration in proving bounds on systems of interacting particles called independent sets; and (c) an extensive semi-autonomous evaluation (Feng et al., 2026a) of 700 open problems on Bloom's Erdos
Conjectures database, including autonomous solutions to four open questions. In order to help the public better understand the developments pertaining to AI and mathematics, we suggest quantifying standard levels of autonomy and novelty of AI-assisted results, as well as propose a novel concept of human-AI interaction cards for transparency. We conclude with reflections on human-AI collaboration in mathematics and share all
prompts as well as model outputs at https://github.com/google-deepmind/superhuman/tree/main/aletheia.
Read the full abstract ▾
All authors Tony Feng, Trieu H. Trinh, Garrett Bingham, Dawsen Hwang, Yuri Chervonyi, Junehyuk Jung, Joonkyung Lee, Carlo Pagano, Sang-hyun Kim, Federico Pasqualotto, Sergei Gukov, Jonathan N. Lee, Junsu Kim, Kaiying Hou, Golnaz Ghiasi, Yi Tay, YaGuang Li, Chenkai Kuang, Yuan Liu, Hanzhao Lin, Evan Zheran Liu, Nigamaa Nayakanti, Xiaomeng Yang, Heng-Tze Cheng, Demis Hassabis, Koray Kavukcuoglu, Quoc V. Le, Thang Luong
Computation and Language · Cross-listed from cs.CR
Omni-Safety under Cross-Modality Conflict: Vulnerabilities, Dynamics Mechanisms and Efficient Alignment
Kun Wang, Zherui Li, Zhenhong Zhou et al. (+8 more)
Omni-modal Large
Language Models (OLLMs) greatly expand LLMs'
multimodal capabilities but also introduce cross-modal safety risks. However, a systematic understanding of vulnerabilities in omni-modal interactions remains lacking. To bridge this gap, we establish a modality-semantics decoupling principle and construct the AdvBench-Omni
dataset, which reveals a significant vulnerability in OLLMs. Mechanistic analysis uncovers a Mid-layer Dissolution phenomenon driven by refusal vector magnitude shrinkage, alongside the existence of a modal-invariant pure refusal direction. Inspired by these insights, we extract a golden refusal vector using Singular Value Decomposition and propose OmniSteer, which utilizes lightweight adapters to modulate intervention intensity adaptively. Extensive experiments show that our method not only increases the Refusal Success Rate against harmful inputs from 69.9% to 91.2%, but also effectively preserves the general capabilities across all modalities. Our code is available at: https://github.com/zhrli324/omni-safety-research.
Read the full abstract ▾
All authors Kun Wang, Zherui Li, Zhenhong Zhou, Yitong Zhang, Yan Mi, Kun Yang, Yiming Zhang, Junhao Dong, Zhongxiang Sun, Qiankun Li, Yang Liu
Computation and Language
Quantum-Audit: Evaluating the Reasoning Limits of LLMs on Quantum Computing
Mohamed Afane, Kayla Laufer, Wenqi Wei et al. (+4 more)
Language models have become practical tools for
quantum computing education and research, from summarizing technical papers to explaining theoretical concepts and answering questions about recent developments in the field. While existing
benchmarks evaluate
quantum code generation and
circuit design, their understanding of
quantum computing concepts has not been systematically measured.
Quantum-Audit addresses this gap with 2,700 questions covering core
quantum computing topics. We evaluate 26 models from leading organizations. Our
benchmark comprises 1,000 expert-written questions, 1,000 questions extracted from research papers using LLMs and validated by experts, plus an additional 700 questions including 350 open-ended questions and 350 questions with false premises to test whether models can correct erroneous assumptions. Human participants scored between 23% and 86%, with experts averaging 74%. Top-performing models exceeded the expert average, with Claude Opus 4.5 reaching 84% accuracy, though top models showed an average 12-point accuracy drop on expert-written questions compared to LLM-generated ones. Performance declined further on advanced topics, dropping to 73% on security questions. Additionally, models frequently accepted and reinforced false premises embedded in questions instead of identifying them, with accuracy below 66% on these critical
reasoning tasks.
Read the full abstract ▾
All authors Mohamed Afane, Kayla Laufer, Wenqi Wei, Ying Mao, Junaid Farooq, Ying Wang, Juntao Chen